Research On Multi-view Learning Methods Based On Fusion Representation

The same object described by data from different sources,structures or angles is called multi-view data,which often has different feature expressions,structures or dimensions.There are various relationships between views,such as correlation,consistency and complementarity.The process of improving task completion performance is multi-view learning by using the relationships between different views and fusion representation through collaborative learning.Fusion representation is a key problem and technical difficulty in multi-view learning.This thesis studies the multi-view learning methods based on fusion representation,explores the influence of view correlation degree on the classification effect based on fusion representation,and studies key technologies such as multi-view subspace learning,subspace clustering,latent space learning,multi-view clustering,etc.The specific research work and main innovations are as follows:(1)Explore the relationship between the degree of multi-view correlation and the classification effect of fusion characterization.From the empirical point of view of data analysis,this thesis adopts open multi-view datasets to calculate the correlation values of the two views based on the maximum information coefficient algorithm and learn to fusion representation and classify prediction result based on two subspace learning models.Based on the above results,explore and analyze the correlation between multi-view features and different view fusion.The taxonomic differences characterized by the degree of view correlation are related to the fusion effect.The experimental results show that the view correlation is strongly related to the view fusion effect;deep learning weakens the influence of view correlation on the fusion effect;and the difference feature is more valuable for improving the fusion effect.The experimental results show that the view correlation is strongly related to the view fusion effect;deep learning weakens the influence of view correlation on the fusion effect;and the difference feature is more valuable for improving the fusion effect.(2)Multi-view subspace learning based on double feedback mechanism.Aiming at the insufficient discrimination expression ability of fusion representation in multi-view subspace learning,a multi-view subspace learning method based on double feedback mechanism is proposed.Considering that the unique differences are implicit in a single view,the dynamic routing mechanism of the capsule network is used to mine the unique characteristics of the single view,and fusion learning is proposed.The minimum difference constraint between the fusion matrix and the network linear transformation output matrix is added to the objective function,which improves the distinguishability of public subspace representation through twostep weight update operations and complete the double feedback learning of feature.The performance of the model is evaluated on the five data sets.The experimental results show that enhancing single-view learning can improve the classification and clustering performance,and the performance improvement effect is significant;the performance of the model is stable in the clustering task;and the performance of the depth subspace learning method in the benchmark algorithm is unstable,and it needs to be further improved.(3)Latent fusion representation based on multi-view latent feature enhancement.Aiming at achieving learning of view consistency and difference in multi-view subspace clustering,a latent space fusion characterization method based on multi-view depth feature enhancement is proposed,which makes full use of the advantages of subspace cluster learning consistency,and adds multi-view implicit space enhancement learning,which can be applied to both classification and clustering task.In this model,the latent space is learned by using the singleview discrimination feature,and then the multi-view self-representation matrix is learned,reconstructing the latent space and self-expression by designing the latent space with the original feature,self-expression and the minimum constraint loss of the latent space.The performance of the model is evaluated on the four data sets.The experimental results show that enhanced view representation improves the clustering and classification performance of the model;the more views there are,the better the clustering performance,and the less reasonableness of clustering;the accuracy of model clustering is higher than that of the benchmark algorithm;when the classification performance of the model varies greatly on a single view,the classification ability of fusion representation is lower.(4)Clustering fusion representation enhancement based on multi-view difference and consistency.In view clustering learning,a clustering fusion enhancement learning method based on multi-view differences and consistency is proposed.Model design is carried out from the difference in view,view consistency,implicit feature enhancement expression of single view and subspace fusion between view.The performance of the model is evaluated on the four data sets,and comprehensive performance evaluation is carried out with the two models proposed in this thesis: the clustering performance of the model is significantly improved,and the relationship between clustering performance and the number of views is found.